AI anxiety: name it to tame it
Adding some structure and perspective to the AI-pocalypse
I recently started this blog after diving into an AI focused role at work to capture my learnings as bite sized, accessible pieces of education on AI. I wanted to make something helpful and optimistic, so that my friends and a handful of internet strangers (hi!) might get ideas on how they could take advantage of opportunities instead of spending their time handwringing about being put out of a job.
But I also want this blog to be a reflection of my actual interests and opinions - and it’s certainly not all sunshine and roses. I personally (as well as most experts) think there’s a lot about AI to be concerned about.
Here’s my take - a generalised sense of anxiety isn’t helpful. It also sucks. What we need is people educated about the specific risks around AI. I don’t think it’s helpful for you and me to worry about robots ‘taking over’. Not only is that pretty speculative, and probably far away in terms of being a real risk, but there are far more specific and real issues we should be aware of happening right now around bias, privacy, and ownership of AI-driven platforms. Not to mention, the energy consumption associated with AI models and the environmental impact (this technology is THIRSTY, y’all).
These are some pretty weighty issues. And, to be honest with you, I’m just a girl with a full time job and a TikTok addiction to keep up. So, I’m just going to start writing a little more about some of my top issues, with my hope being that adding some structure around how AI could shape our lives into the future makes you feel a little more prepared, and a little less anxious. Maybe, it’ll even plant some seeds around how you could contribute. I know that’s one thing I am deeply reflecting on, myself.
I hope you find it useful!
Char
Xx
On AI and Bias
A dear friend of mine who is a total rockstar in the domestic violence prevention space asked me recently to write about bias in AI. It got me thinking about how the ‘black box’ nature of these systems makes it really hard for people to even pinpoint when bias is happening and what it looks like, let alone challenge it.
You might have heard a lot talk about how ‘AI platforms are racist’ or that they lead to unfair outcomes when they’re used by governments, or medicine, or in hiring.
There are a lot of systems out there in the wild that absolutely are racist, or misogynistic, or discriminatory in some way. And these systems have actually been impacting people’s lives for years. So let’s dive into why.
Garbage in, garbage out
At the core of it is a simple principle: garbage in, garbage out.
Think of a small child. They learn about the world through the data that they take in through their five senses. A lot of the information they gain about the world comes from their parents, and, well, if they have parents who teach them racist information, they’re going to parrot back racist information.
AI models get trained on data, and if it’s not good data, you won’t get good outputs.
AI models need data, and lots of it. LLMs like ChatGPT are called large language models because they’re trained on basically the entire internet (a much larger data set than previous iterations).
And collecting data is, well… expensive. Early image recognition model ImageNet used millions of images manually labelled by people who were had to be paid for their time. Because collecting fresh data for a specific use case is so expensive, it means that it became commonplace to just reuse data you already had lying around for other purposes.
The trouble with using data you already have lying around is that you might not always be measuring quite the right thing. For example, a model that’s meant to predict the likelihood of a person’s risk of committing a crime might *actually* be measuring their probability of being arrested, since that’s the data they used - which may have police officers’ bias baked in.
Making predictions about people is hard
I mentioned there are already AI systems out in the wild that are extremely biased, creating large scale harm. One of the most notorious is the COMPAS system used by judges in the United States to make decisions around bail and sentencing.
The system is meant to flag offenders deemed ‘high risk’ based on grouping people together with similar characteristics. It generates a high number of false positives for black people, and false negatives for white people. This happens because the data used to train the model includes information about arrest records, employment and housing status - meaning in practice that if you happen to be poor and black, you could be denied bail and wrongfully imprisoned for years. In the book AI Snake Oil, Narayan and Kapoor explain that despite the hundreds of data points used in COMPAS, the predictions on whether someone will re-offend are so inaccurate that it’s not much better than a coin toss, and that it would be almost as good to just use a single data point: has the person offended in the past?
Over in the UK they use a similar system called Oasys, and governments around the world have for years been promoting the idea of unbiased AI justice tools. In theory, these kinds of tools could remove subconscious bias by individual judges or case workers - but in practice, there’s a real risk that the bias gets baked in and even magnified. A lot of developers will caveat that the outputs of these systems require a human review due to the risk of inaccurate results, but in reality, workers are often uncomfortable with challenging the outputs.
One of the biggest challenges around predictive AI when applied to people is that people are hard to predict. Chance events and individual differences mean that blanket tools that calculate our chances of some behaviour are usually flawed.
AI literacy is critical
In my view, the key piece here is that currently, most people just don’t understand how these systems work well enough to judge whether they are biased or not. AI and machine learning tools tend to be these backend-y, mysterious things that we trust to just spit out correct information, even when it’s making consequential decisions about people’s lives. What’s more, the publishers of these systems often don’t release the details around what features were used to train the models, making it difficult for outsiders to understand anything about how it works.
Developers and adopters of AI tools have a responsibility to prevent these systems from causing harm. This includes assessing bias and rigorously testing accuracy, including after the systems are in place to avoid unintended side effects. In an ideal world, AI tools could actually be used to reduce the bias that goes into all kinds of decision making - but the systems need to be assessed much more thoroughly, likely with a lot more transparency around how they work.
When we the public aren’t informed and educated, it’s really hard for us to push back on inappropriate and harmful systems when developers and adopters of this tech haven’t implemented responsible solutions.
Hopefully regulation catches up to help prevent harm, but making sure you understand the basics is also super important.
If you want to read more, I can highly recommend the books AI Snake Oil by Arvind Narayan and Sayash Kapoor, and Nexus from Yuval Noah Harari.
For more bite sized content, Dr. Avriel Epps is an amazing creator to follow on TikTok or Instagram - her handle is @kingavriel. She talks about bias in AI in a super accessible format and is a researcher at Cornell.